← Paper Digest

Finetuning VLAs for Dexterous Manipulation: The Right Checkpoint Is All You Need

Artur Kuramshin, Charlotte Cloutier, Michael Przystupa, Kirsty Ellis, Mohammad S. Akhter, Glen Berseth

November 2026

TL;DR They fine-tune π0.5 for a 16-DoF LEAP hand with five recipes and steer each checkpoint with diffusion-steering RL in simulation. Training only the action expert steers to 95-100% success on training placements in both tasks, and SFT rollout success does not predict which checkpoint steers best.

What they do

VLA pretraining data comes mostly from parallel-jaw grippers, so a multi-fingered hand needs fine-tuning on a few dozen demonstrations, often followed by RL. The authors ask which supervised fine-tuning (SFT) recipe and which checkpoint make the best starting point for diffusion-steering RL (DSRL), which keeps the VLA frozen and learns only the noise fed to its action expert. They test π0.5 on a 16-DoF LEAP hand in two RoboCasa simulation tasks.

How it works

  • Five recipes. Expert-only (whole VLM frozen), Frozen-LM (also trains the vision encoder and projector), Full-FT, LoRA on the language model and expert, and LoRA-LM (LoRA on the language model only).
  • Data. 33 teleoperated demos for ScrewLightbulb (xArm6) and 56 for CoffeePressButton (Franka), collected in simulation with a Quest 3 headset and a Rokoko glove.
  • Steering. For four checkpoints per recipe and task, a SAC actor picks one 32-D noise vector per action chunk from a 64×64 image and proprioception; 500k environment steps, three seeds.
  • Evaluation. Two training placements and two unseen extrapolation placements per task, scored before and after RL.

Results

  • Expert-only is the only recipe that steers to 95-100% on training placements at every checkpoint. At 8k steps on ScrewLightbulb it reaches 100% on training and 88% on extrapolation placements; every other recipe is at or below 33%.
  • SFT success has no significant positive correlation with post-DSRL success on ScrewLightbulb and is anti-correlated on CoffeePressButton. There, LoRA at 18k has the highest repeated-noise SFT success (76%) and the lowest post-DSRL success.
  • Recipes that update the language model do worst, with or without LoRA. Frozen-LM matches Expert-only early but collapses later (0% on ScrewLightbulb at 8k and 16k).

Limitations

One SFT run per recipe, so SFT variance is not measured. Simulation only, and steering exploits it: early ScrewLightbulb checkpoints turn the bulb with the palm instead of grasping it. No offline metric yet predicts which checkpoint will steer well.

Figures

From Figure 1: The two tasks. ScrewLightbulb varies the lamp pose on the counter and CoffeePressButton the lateral of...
From Figure 1: The two tasks. ScrewLightbulb varies the lamp pose on the counter and CoffeePressButton the lateral offset of the coffee machine; policies are scored on training placements and on unseen extrapolation placements.
From Figure 2: Success before (dashed) and after (solid) steering for each recipe, by SFT step, on the training and e...
From Figure 2: Success before (dashed) and after (solid) steering for each recipe, by SFT step, on the training and extrapolated placements of both tasks.
From Figure 5: Steered Expert-only rollouts on ScrewLightbulb. The 2k-step checkpoint turns the bulb with the palm wi...
From Figure 5: Steered Expert-only rollouts on ScrewLightbulb. The 2k-step checkpoint turns the bulb with the palm without a stable grasp; the 16k-step checkpoint grasps and twists as in the demonstrations.
Abstract

Vision-Language-Action models have demonstrated strong manipulation capabilities. A limitation of VLA models is that most manipulation datasets focus on parallel-jaw gripper systems, making transfer to multi-finger hands challenging, particularly as teleoperation datasets are harder to collect with multi-fingered robots. In this work, we study how to best adapt VLAs to dexterous hands by pairing supervised fine-tuning (SFT) with diffusion-steering reinforcement learning (DSRL). Using $π_{0.5}$ and a LEAP hand, we evaluate five fine-tuning recipes across two simulation tasks using only a handful of demonstrations. Our empirical analysis yields three key insights: (1) SFT success is poorly correlated with post-DSRL success, meaning the best SFT checkpoint is rarely the most effective to steer; (2) freezing the vision-language model (VLM) and tuning only the action expert produces the most steerable policy for both training and out-of-distribution placements; and (3) low-rank adaptation (LoRA) is the least steerable approach. Consequently, we recommend an adaptation recipe that freezes the VLM, fully trains the action expert, and eschews SFT rollout success as a checkpoint selection metric.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.